Causal Explanation Framework for AI Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Constructing causal explanations for machine learning models is challenging due to the difficulty in changing data aspects without leaving the data distribution, which is essential for explaining how input data modifications affect classifier outputs.
Innovation Solution
A learning framework that uses a generative model and information-theoretic measures to identify causal influences, allowing for the generation of causal explanations by filtering data into a disentangled low-dimensional representation and determining generative mappings between latent factors and classifier outputs, without requiring labeled attributes or knowledge of causal structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If data aspects are changed to explain classifier output, then explanatory power is improved, but the modified data may leave the data distribution making explanations meaningless
Solution Approach 1:
The patent introduces latent factors as intermediary variables that mediate between the original data and the classifier output. These latent factors serve as a bridge that allows explanation of causal relationships while maintaining validity within the data distribution, as they are learned representations that capture essential variations without requiring out-of-distribution modifications
2Reliability
If a generative model is used to represent data distribution, then data distribution fidelity is improved, but model complexity increases
Solution Approach 1:
The patent extracts the essential data distribution characteristics into a low-dimensional latent space, separating the core distributional properties from the full data complexity. This extraction allows the generative model to capture data distribution fidelity with reduced complexity by focusing on the most important latent factors rather than modeling the entire high-dimensional data space
3Loss of information
If latent factors are optimized to have large causal influence on classifier output, then explanation quality is improved, but the factors may not represent the true data distribution
Solution Approach 1:
The patent segments the latent factors into two distinct groups: those optimized for causal influence on the classifier output (explanatory factors) and those optimized to represent the data distribution (distributional factors). This segmentation allows each group to specialize in its respective function without compromising the other, enabling both high explanation quality and accurate data distribution representation
Data Source
AI summary
Methods, non-transitory computer readable media, and causal explanation computing apparatus that assists with generating and providing causal explanation of artificial intelligence models includes obtaining a dataset as an input for an artificial intelligence model, wherein the obtained dataset is filtered to a disentangled low-dimensional representation. Next, a plurality of first factors from the disentangled low-dimensional representation of the obtained data that affect an output of the artificial intelligence model is identified. Further, a generative mapping from the disentangled low-dimensional representation between the identified plurality of first factors and the output of the artificial intelligence model, using causal reasoning is determined. An explanation data is generated using the determined generative mapping, wherein the generated explanation data provides a description of an operation leading to the output of the artificial intelligence model using the identified plurality of first factors. The generated explanation data is provided via a graphical user.